PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentIR: Reasoning-Aware Retrieval for Deep Research Agents

Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, Victor Zhong

arXiv:2603.04384v3cs.CL

TL;DR

Deep Research agents generate reasoning traces that contain search intent and context unavailable to conventional retrievers. The paper jointly embeds these traces with queries and synthesizes training data with DR-Synth, producing AgentIR-4B, which reaches 68% BrowseComp-Plus accuracy with Tongyi-DeepResearch versus 50% for conventional embeddings and 37% for BM25.

  • Problem

    Deep Research agents expose explicit reasoning before searches, but existing retrievers ignore this signal and lack training data tailored to their multi-turn sub-queries.

  • Method

    Reasoning-Aware Retrieval jointly embeds agent reasoning traces with queries, while DR-Synth creates sub-query–relevance training pairs from standard QA datasets.

  • Results

    68% accuracy on BrowseComp-Plus with Tongyi-DeepResearch exceeds 50% for conventional embedding models twice its size and 37% for BM25.

  • Takeaways & Limitations

    Reasoning-Aware Retrieval and DR-Synth are independently effective, and their combination yields consistent gains across multiple LLM agents.

Abstract

from arXiv · show

Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent's reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68\% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50\% with conventional embedding models twice its size, and 37\% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.

1 Introduction

Deep Research agents expose reasoning traces that clarify search intent, but conventional retrievers ignore this signal. AgentIR combines reasoning-aware retrieval with DR-Synth training data and achieves substantial accuracy, efficiency, and cross-agent gains.

  • Reasoning-Aware Retrieval jointly embeds each issued query with the agent’s natural-language reasoning trace to use intent and contextual information.The traces include reflections on prior results, unresolved gaps, and hypotheses about promising search targets.
  • 68% end-to-end accuracy on BrowseComp-Plus with Tongyi-DeepResearch exceeds 52% for a conventional embedding model twice its size and 37% for BM25.AgentIR-4B also outperforms computationally intensive LLM-based reranking by 10% absolute.
  • AgentIR-4B reduces search steps without additional inference overhead because the reasoning traces are already generated by the agent.
  • Without additional training, AgentIR-4B generalizes its gains across agents with different reasoning patterns, including gpt-oss-120B and GLM-4.7.
  • Reasoning traces help retrieval by summarizing earlier findings while implicitly filtering outdated or incorrect information.
  • DR-Synth constructs agent sub-query–relevance pairs from standard QA datasets, addressing the lack of training data for Deep Research retrieval.

2 Related Work

Deep Research retrieval differs from conventional retrieval because agents interleave explicit reasoning with multi-turn search. Related work addresses reasoning-intensive retrieval and ambiguous queries, but this paper targets collaborative retrieval using the agent’s evolving state.

  • Deep Research agents autonomously conduct multi-turn searches, often executing extensive retrieval while solving complex problems.
  • Reasoning-Aware Retrieval focuses on collaborative retrieval with an agent over multiple turns rather than requiring the retriever to solve the task in one turn.
  • Unlike human-query methods that add instructions, clarification, or hypothetical context, this approach exploits reasoning traces already generated during agent interaction.

3 Methodology

AgentIR models Deep Research as a multi-turn reasoning-and-search loop, then jointly embeds each reasoning trace with its query. DR-Synth converts standard QA data into sub-query-level supervision using agent rollouts and oracle reranking.

  • Deep Research: Deep Research agents generate reasoning traces and search actions conditioned on interaction history, receiving retrieval feedback across multiple turns.The action is either a search call or a final answer, and search queries are denoted by q_t.
  • Reasoning-Aware Retrieval: Conventional retrieval embeds only q_t, whereas Reasoning-Aware Retrieval embeds the current reasoning trace and query together as R(τ_t, q_t).The reasoning clarifies task intent, incorporates prior findings, and supplies hypothetical search targets without an additional query-expansion call.
  • Training objective: The resulting contrastive training instances concatenate τ_t and q_t, with cosine similarity optimized using temperature T = 0.01.The concatenated input is represented as [τ_t, q_t].
  • DR-Synth: DR-Synth addresses the lack of sub-query supervision by generating reasoning-query instances from standard (Q, A, P) datasets through agent rollouts.Global positive documents do not directly identify documents relevant to each local sub-query, motivating turn-specific supervision.
  • DR-Synth: Oracle reranking labels turn-specific positives and hard negatives by reranking retrieved candidates with the current query, global question, and true answer.The top-ranked document becomes d+_t, while the bottom seven become hard negatives; successful rollouts are retained for training.

4 Experiments

AgentIR-4B is evaluated with three open-weight Deep Research agents on BrowseComp-Plus against query-only, query-expansion, and reranking baselines. It achieves the best reported performance while also reducing search effort and avoiding reranking overhead.

  • Overall results: AgentIR-4B achieves the best performance across all evaluated agents on BrowseComp-Plus, outperforming prior baselines and concurrent work.The evaluation reports end-to-end accuracy, retrieved-document recall, and search-call counts for three open-weight agents.
  • Results: 66.27% accuracy for Tongyi-DR is 17.60% above Qwen3-Embed-4B and approximately 15% above Qwen3-Embed-8B, despite the latter being twice as large.The 17.60% gain is comparable to the reported 14.69% gain from BM25 to Qwen3-Embed-4B.
  • Efficiency: AgentIR-4B reduces Tongyi-DR search calls from 32.92 with BM25 to 25.91 and exceeds Qwen3-Embed-4B plus LLM reranking by approximately 10% absolute accuracy.AgentIR-4B performs no reranking, whereas the comparison reranks the top 20 results with Qwen3-8B.
  • Baselines: AgentIR-4B surpasses ReasonIR-8B by approximately 15% accuracy, while HyDE-like query expansion is reported ineffective because its hypotheses can misinterpret the query and hallucinate.Direct access to the agent’s reasoning produces more grounded hypotheses than context-free expansion.
  • Generalization: AgentIR-4B generalizes zero-shot from WebShaper-trained trajectories to distinct agents, including gpt-oss-120B and GLM-4.7, without additional fine-tuning.The model also transfers to a Tongyi-DR setting that includes the visit tool.

5 Analysis

Ablations show that reasoning traces and DR-Synth training contribute independently, with their combination strongest across agents. Further analyses find that current reasoning summarizes useful history while filtering noise, making additional prior turns largely redundant or harmful.

  • Component ablation: 66.27% accuracy results when DR-Synth training and reasoning-aware inputs are combined, versus 59.4% with training alone and 55.54% with reasoning alone on Tongyi-DR.Both components are independently effective, and the same pattern holds across all three agents.
  • Alternative signals: AgentIR-4B outperforms prior queries and the global question, while adding prior queries with reasonings improves over prior queries alone but remains below the current reasoning-query input.The alternatives vary the trajectory transformation used as the retriever input.
  • Alternative signals: Embedding prior documents introduces noisy context: 11.45% of Tongyi-DR runs have zero recall and average 37.46 search turns, versus 27.18 overall.The analysis attributes the prolonged failures to irrelevant searches being propagated into later queries.
  • Effect of prior reasonings: Adding prior reasoning turns beyond the current one does not improve accuracy, while the current reasoning alone covers more than 40% of all past clues with diminishing returns as k increases.Because each reasoning is conditioned on the full history, it often summarizes earlier findings already needed for retrieval.
  • Effect of prior reasonings: Earlier reasonings add substantially more incorrect than correct claims, whereas current reasoning omits failed candidates and filters outdated or incorrect guesses.This implicitly curated history is presented as a cleaner retrieval signal than naively embedding the full reasoning history.

6 Conclusion

The paper introduces Reasoning-Aware Retrieval and DR-Synth for Deep Research agents, combining them in AgentIR-4B. It finds that curated reasoning context improves retrieval and motivates context engineering for retrievers.

  • Reasoning-Aware Retrieval and DR-Synth are independently effective, while their combination in AgentIR-4B substantially outperforms existing retrievers across multiple LLM agents.
  • Reasoning traces ground each search in historical context and naturally filter incorrect prior hypotheses, whereas embedding more uncurated history underperforms.
  • Agent-optimized context may be leveraged directly by retrievers, motivating principled context curation without additional computational cost.
  • The work targets a future in which autonomous agents increasingly become primary consumers of search and humans consume their results.

B AgentIR-4B Training Details

AgentIR-4B is trained by fine-tuning Qwen3-Embedding-4B on DR-Synth instances that pair reasoning-aware queries with positive documents. The training corpus combines scraped positives, search-derived hard negatives, and random FineWeb negatives.

  • AgentIR-4B fine-tunes Qwen3-Embedding-4B with LoRA and contrastive learning on inputs concatenating reasoning trace τt with query qt.
  • Training uses learning rate 1e-4, batch size 4, maximum document length 4096, maximum query length 8192, and two epochs on a single H100.
  • DR-Synth rollouts use Tongyi-DR, Qwen3-Embedding-8B retrieval, and GLM-4.6 oracle reranking, returning five ranked documents to continue each rollout.
  • The training corpus includes scraped positive URLs, GPT-4o-decomposed sub-query hard negatives retrieved through SerpAPI, and one million random FineWeb negatives.
  • 5,238 training instances are obtained from 250 correctly answered WebShaper trajectories after rollout over 1,146,942 deduplicated documents.

C Prompts for Alternative Sources of Retrieval Signals

Alternative retrieval-signal variants are trained with the same procedure but replace rollout-derived reasoning-query inputs with inputs constructed from selected history functions and variant-specific templates.

  • Each alternative variant uses the same training procedure, changing only the retriever input constructed from f(Ht).
  • The Prior Queries variant concatenates q1 through qt using its dedicated prompt template and retains the same positive documents d+t.
  • The Prior Queries & Reasonings and Prior Queries & Reasonings & Docs variants use separate prompt templates shown in Figures 8 and 9.

D Prompt for Adding Prior Reasonings

The prior-reasonings study trains checkpoints using different numbers of recent turns, while prompt formatting preserves a normalized turn index and frames retrieval as using browsing history with the current query.

  • D Prompt for Adding Prior Reasonings: The study trains separate checkpoints for k ∈ 1, 2, 5, 9, 17, and all prior turns; k = 1 is AgentIR-4B and requires no additional training.
  • D Prompt for Adding Prior Reasonings: Each input includes the selected sequence of reasoning-query pairs from j = max(1, t − k + 1) through t.
  • D Prompt for Adding Prior Reasonings: The prompt uses the Figure 8 template but includes only the most recent k turns rather than all prior turns.
  • D Prompt for Adding Prior Reasonings: Although selected turns retain their global positions, the prompt relabels them sequentially from 1 to k.
  • D Prompt for Adding Prior Reasonings: The retrieval task combines a user's prior browsing history with the current web-search query to retrieve relevant answer passages.
  • D Prompt for Adding Prior Reasonings: Figures 7 and 8 provide prompts for Prior Queries and Prior Queries & Reasonings ablations, respectively.
  • E Prompts for Atomic Clues: Figures 10 and 11 show prompts for decomposing reasonings into atomic clues and assigning those clues to documents.

F HyDE Example

Figure 12 presents a full example of Reason-Rewriter’s HyDE-style expansion for the query “backroom studio early 2010s euphoric”.

  • Figure 12 shows Reason-Rewriter’s complete HyDE-style expansion for “backroom studio early 2010s euphoric”.

G Prompt for Noise Analysis

The paper uses staged prompts to decompose reasoning traces into clues, assign clues to reasoning text, and evaluate intermediate answers against evidence documents.

  • The noise-analysis procedure extracts intermediate answers from evidence documents before judging claims, avoiding errors caused by seeing only final answers.
  • The ablation prompt evaluates the role of prior queries, reasonings, and documents, while an example generated document illustrates the broader analysis setup.
  • The workflow decomposes reasoning traces into short, independent clues that capture the overall solving process without redundancy.
  • The supporting figures document prompts for clue decomposition, clue assignment, multi-hop answer extraction, and correct-versus-incorrect claim counting.
  • A separate prompt identifies which clue numbers are mentioned or related to each text paragraph.
Loading 2603.04384v3…